The Claude Opus 5.5 system card reveals an opaque evaluation process that raises fresh questions about transparency in AI safety testing
Anthropic’s latest system card for Claude Opus 5.5 contains a detail that deserves more attention than it’s getting: the independent nonprofit METR, tasked with evaluating whether the model accelerates AI research to dangerous levels, relied on an undisclosed source of information to reach its conclusions. The evidence behind those conclusions can’t be shared publicly.
What the system card actually says
The Claude Opus 5.5 System Card, released on September 22, 2026, documents the pre-deployment evaluations Anthropic conducted before releasing its latest Opus-class model. Among the key findings: METR’s external testing did not identify a sustained doubling in the pace of AI development compared to Anthropic’s own internal assessments.
But the methodology behind that conclusion is where things get murky. The METR team operated with restricted access, and their report was described as “highly experimental and preliminary.” The system card also notes that Anthropic itself utilized another undisclosed source of information in its evaluations, a detail buried in a document most people will never read in full.
What Opus 5.5 actually brings to the table
The Opus 5.5 model itself represents a meaningful step forward for Anthropic’s product line. The system card documents improvements across several domains, including agentic coding capabilities, more effective computer usage, and stronger performance on long-horizon professional tasks.
What this means for AI development accountability
The real test will come when METR or a similar evaluator does find evidence of dangerous capability acceleration. Whether undisclosed sources and restricted access would be acceptable in that scenario, when the stakes are no longer theoretical, is a question the current framework hasn’t answered.
